Industries · 7 min read
Diabetes device inspection with a VLM and one detector across four device shapes
Four device shapes on one line: a question judges the faults nobody wrote a class for, a detector owns the count, and a new device generation is labeled first.
Rob Hickey · Sep 30, 2026
Operations · 6 min read
How to test computer vision model robustness before the weather does it for you
A vest model trained in June daylight meets October rain. Blur, darken, shift and compress the held-out frames first, and read recall per perturbation.
Rob Hickey · Sep 30, 2026
Industries · 6 min read
Self-driving car perception and the white trailer against a bright sky
A model cannot find what its labels never showed it, the miss looks like nothing on the frame, and doubted fleet frames are how the training set catches up.
Rob Hickey · Sep 30, 2026
Operations · 6 min read
Underspecification in machine learning, and why two models with the same score do not both survive the second plant
Fifty training runs, one recipe, one validation score. The test set could not tell them apart, and the second plant's cameras can.
Rob Hickey · Sep 30, 2026
Industries · 7 min read
Wildfire smoke detection from a fixed ridge camera at dawn
Smoke is a class with no edges, cloud and valley mist are the false alarms to label, and a camera that never moves is what makes the first thin column visible.
Rob Hickey · Sep 30, 2026
Industries · 7 min read
Zero-shot pose estimation for a collaborative robot, and what makes it hold
A generalist pose model reads a worker's reach on day one, keypoints are smoothed across sampled frames, and the site's own reviewed frames make it hold.
Rob Hickey · Sep 30, 2026
Industries · 7 min read
Aerial fire detection from a drone patrol, smoke before the flame reaches the line
On a right-of-way patrol, smoke is a few dozen grey pixels that look like haze. Two boxed classes, an alert with a cooldown, and the bad-weather days kept.
Rob Hickey · Sep 29, 2026
Industries · 6 min read
Automated sorting with computer vision, from the camera over the conveyor to the diverter
A box on every apple, a grade from the box, and an air jet that acts on it before the belt moves on. The new cultivar is when the model needs the graders again.
Rob Hickey · Sep 29, 2026
Industries · 6 min read
Infrastructure asset management with computer vision from a truck at road speed
Signs, guardrails and poles pass the truck camera in a fraction of a second. The model finds them, the register says what should be there, a crew gets the gap.
Rob Hickey · Sep 29, 2026
Operations · 7 min read
Benchmark the model on your own cameras before you believe a score
Two candidate models, two published scores, and a packaging line that only cares which one finds the torn label. The held-out week from your cameras decides.
Rob Hickey · Sep 28, 2026
Computer vision · 6 min read
Reading a training run before trusting the model
The loss curve says the run finished. The gap to validation, the recall on the crack class and the frames behind the score say whether the line can trust it.
Rob Hickey · Sep 28, 2026
Operations · 6 min read
The class that fails behind a good average, read off the confusion matrix
A shelf model with a healthy mAP and a dairy case full of unreported gaps. The matrix names the cell the failing class fell into, and the cell names the fix.
Rob Hickey · Sep 28, 2026
Operations · 7 min read
Precision, recall and the mistakes each one lets through
A shelf model that cries wolf and an inspection model that waves the crack past are wrong in opposite ways. The threshold is the dial between them.
Rob Hickey · Sep 28, 2026
Labeling · 6 min read
Out-of-scope detection, what the model does when it sees something it was never taught
A screw detector on the phone line meets an operator's hand and calls it a screw. Null frames, a class for confusing things, a test set from the real camera.
Rob Hickey · Sep 28, 2026
Computer vision · 6 min read
Real-time object detection models, transformer or convolutional, and which size for the runner
A DETR-style detector drops the anchors and the suppression step. The size to run is decided by the box beside the camera, which the leaderboard cannot see.
Rob Hickey · Sep 28, 2026
Labeling · 7 min read
What active learning is, and why it labels the frames that change the model
Fifty thousand frames of good panels and nine splits. Active learning sends the frames the model doubts to a person first, and the cycle runs on the live line.
Rob Hickey · Sep 28, 2026
Operations · 7 min read
Active learning for computer vision on a line camera that never stops
The weld camera runs three shifts. The model returns the frames it doubts, the inspector corrects them, and past the threshold a new version trains and ships.
Rob Hickey · Sep 27, 2026
Operations · 7 min read
How to improve computer vision model accuracy by starting with one defect class
A coffee bag line wanted eight defect classes and got none. Torn bag first, frames from the station camera, a written guideline, then the loop adds the rest.
Rob Hickey · Sep 27, 2026
Operations · 6 min read
Human-in-the-loop computer vision, the mistakes the model is sure about and the inspector who catches them
A quality inspector flags a wrong detection in seconds, the confidently wrong frames no threshold would catch, and the correction rate as the signal.
Rob Hickey · Sep 27, 2026
Operations · 6 min read
Monitoring a computer vision model in production, and what to do when the corrections start climbing
A shelf model meets the Christmas displays: the correction rate as the signal, doubted frames first in the queue, a new version compared on the same frames.
Rob Hickey · Sep 27, 2026
Operations · 8 min read
Monitoring data drift in computer vision, and the one signal that catches both kinds
A container yard camera at night is data drift. A damage policy that now counts dents is concept drift. Only the correction rate catches both.
Rob Hickey · Sep 27, 2026
Operations · 7 min read
Production deployment is the halfway point of a computer vision model's life
A packaging line adds a flavour and a store resets for autumn. The model that scored well in evaluation meets both on a Monday, and the second half begins.
Rob Hickey · Sep 27, 2026
Industries · 6 min read
AI crop analysis in the greenhouse, catching tomato disease before it spreads down the row
Lesions boxed by disease with a healthy class, a question of the footage about how far a patch spread since last week, and a model that turns with the season.
Rob Hickey · Sep 26, 2026
Industries · 6 min read
Juice box quality inspection at the straw applicator, with a tag on every carton
Acceptable, missing, bent or pierced on every carton from the line's own frames, and why a nudged camera mount is the failure a classifier shows first.
Rob Hickey · Sep 26, 2026
Industries · 7 min read
Predictive maintenance with computer vision on the bearing that is about to seize
A camera on the bearing housing, the belt and the pipe joint sees the weep, the fray and the loose bolt weeks before the stop, and sends the frame.
Rob Hickey · Sep 26, 2026
Industries · 6 min read
Robotic pick and pack with a bounding box and a mask on every item, and why the arm still misses
A box finds the item, a mask finds where the cup should land, foil packaging punches holes in the depth map, and a nudged camera mount shows up as failed picks.
Rob Hickey · Sep 26, 2026
Industries · 8 min read
Computer vision in steel manufacturing, from the ladle shell to the coil
A thermal camera on the ladle stand sees the lining fail from the outside, and a replaced camera is the day the hotspot map has to be redrawn.
Rob Hickey · Sep 25, 2026
Industries · 7 min read
Contact lens defect inspection with pass, review and fail as three answers
A dark-field camera turns a crack in a clear lens into a bright line. The lens the model cannot decide goes to a person, and that middle bucket is the product.
Rob Hickey · Sep 25, 2026
Industries · 7 min read
How to detect small defects with computer vision when three resizes erase them
A hairline crack on a blade and a micro-scratch on a display vanish before the first layer sees them. Tiling keeps their pixels; the labeler sees them first.
Rob Hickey · Sep 25, 2026
Industries · 6 min read
Drone structural damage detection with cracks and spalling confirmed by the engineer
Masks for cracks and spalling from a close drone pass, so area compares pass to pass. The engineer confirms each flag, and a re-planned flight moves every mask.
Rob Hickey · Sep 25, 2026
Industries · 7 min read
Fall detection with pose estimation on the warehouse cameras a plant already has
A worker down behind the racking at 2 am, a torso line that drops between samples, and an alert with the frame attached that reaches someone who can walk there.
Rob Hickey · Sep 25, 2026
Industries · 6 min read
Flaw detection with computer vision for the flaws an inspector stops seeing at hour eight
Surface, structural and material flaws on one welded bracket, why inspectors disagree by hour eight, and the review loop that holds the 7 am standard.
Rob Hickey · Sep 25, 2026
Industries · 6 min read
AI visual inspection as the nondestructive testing step a camera can take over
Visual testing is the first NDT gate, its acceptance criteria are already written, and a camera can apply them to every weld instead of one in twenty.
Rob Hickey · Sep 24, 2026
Industries · 6 min read
Computer vision pill inspection catches one damaged tablet in a thousand
A camera over the tablet conveyor before the blister packer, recall set on the defect side, and thresholds re-tuned when a new batch moves the base rate.
Rob Hickey · Sep 24, 2026
Industries · 6 min read
Machine vision or computer vision on the line, and when each breaks
A threshold rule that died with the old fluorescent tubes, a template that broke on the new variant, and the model that is corrected rather than re-engineered.
Rob Hickey · Sep 24, 2026
Industries · 6 min read
On-shelf availability monitoring with the aisle camera a store already has
Empty facings boxed on the bread bay, glare and black packaging as the false alarms, a cooldown rule to the restocker, alert volume watched against itself.
Rob Hickey · Sep 24, 2026
Labeling · 6 min read
Writing a class ontology for defect labeling before the first box
On a weld line, the boundary, occlusion and inclusion rules come before the first box. When the tolerance tightens, the old labels are wrong.
Rob Hickey · Sep 23, 2026
Industries · 8 min read
Reading analog gauges with a camera, without replacing them
A camera on the dial turns a clipboard round into a stream of readings, and a nudged bracket is what quietly breaks it.
Rob Hickey · Sep 23, 2026
Industries · 6 min read
Reducing manufacturing scrap with a camera on the line, because scrap is a timing problem
A torch drifts at 9 am and the shift finds out at 2 pm. Scrap scales with that lag, and the camera flags the first bad bead rather than the hundredth.
Rob Hickey · Sep 23, 2026
Industries · 6 min read
Red zone monitoring with a camera over the forklift lane
Person and vehicle classes trained on the site's own frames, the zone as a polygon, dwell as the rule, and a nudged camera as what silently moves the zone.
Rob Hickey · Sep 23, 2026
Industries · 7 min read
Transmission line inspection with object detection from a survey flight
Damaged strands, broken insulators, foreign objects and vegetation from one flight, tiled so a strand keeps its pixels, with recall set per hazard.
Rob Hickey · Sep 23, 2026
Industries · 6 min read
Visual anomaly detection on the line, catching the defect you never labeled
Learn normal from clean frames, score every fillet against it, route the strange ones to a person, and turn what they name into classes over time.
Rob Hickey · Sep 23, 2026
Computer vision · 7 min read
How to catch model drift before your customers do
A cap-check model passed capless bottles for six weeks and nobody knew until a retailer rang. The signal was there the whole time, in the review queue.
Rob Hickey · Jul 14, 2026
Case notes · 7 min read
Corrosion detection on transmission insulators, from the inspection pass to a pin you can name
A drone pass along an extra-high-voltage line shows every shed and pin. The frame with early rust looks like the one before it; tracking makes the call stick.
Rob Hickey · Jul 2, 2026
Edge · 6 min read
Why your best vision model should live at the edge, next to the camera
A substation at the end of a long feeder has a thermal camera, a cellular uplink that drops in rain, and a model that keeps watching when the link does not.
Rob Hickey · May 22, 2026